]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blobdiff - fs/btrfs/async-thread.c
btrfs: Add threshold workqueue based on kernel workqueue
[mirror_ubuntu-bionic-kernel.git] / fs / btrfs / async-thread.c
index 8aec751fa464c126a7fd80e21b526692896d98c9..977bce2ec887652c0101d60aeeae37d0e20462d8 100644 (file)
@@ -1,5 +1,6 @@
 /*
  * Copyright (C) 2007 Oracle.  All rights reserved.
+ * Copyright (C) 2014 Fujitsu.  All rights reserved.
  *
  * This program is free software; you can redistribute it and/or
  * modify it under the terms of the GNU General Public
@@ -21,6 +22,7 @@
 #include <linux/list.h>
 #include <linux/spinlock.h>
 #include <linux/freezer.h>
+#include <linux/workqueue.h>
 #include "async-thread.h"
 
 #define WORK_QUEUED_BIT 0
@@ -28,6 +30,9 @@
 #define WORK_ORDER_DONE_BIT 2
 #define WORK_HIGH_PRIO_BIT 3
 
+#define NO_THRESHOLD (-1)
+#define DFT_THRESHOLD (32)
+
 /*
  * container for the kthread task pointer and the list of pending work
  * One of these is allocated per thread.
@@ -262,18 +267,19 @@ static struct btrfs_work *get_next_work(struct btrfs_worker_thread *worker,
        struct btrfs_work *work = NULL;
        struct list_head *cur = NULL;
 
-       if (!list_empty(prio_head))
+       if (!list_empty(prio_head)) {
                cur = prio_head->next;
+               goto out;
+       }
 
        smp_mb();
        if (!list_empty(&worker->prio_pending))
                goto refill;
 
-       if (!list_empty(head))
+       if (!list_empty(head)) {
                cur = head->next;
-
-       if (cur)
                goto out;
+       }
 
 refill:
        spin_lock_irq(&worker->lock);
@@ -495,6 +501,7 @@ static int __btrfs_start_workers(struct btrfs_workers *workers)
        spin_lock_irq(&workers->lock);
        if (workers->stopping) {
                spin_unlock_irq(&workers->lock);
+               ret = -EINVAL;
                goto fail_kthread;
        }
        list_add_tail(&worker->worker_list, &workers->idle_list);
@@ -725,3 +732,293 @@ void btrfs_queue_worker(struct btrfs_workers *workers, struct btrfs_work *work)
                wake_up_process(worker->task);
        spin_unlock_irqrestore(&worker->lock, flags);
 }
+
+struct __btrfs_workqueue_struct {
+       struct workqueue_struct *normal_wq;
+       /* List head pointing to ordered work list */
+       struct list_head ordered_list;
+
+       /* Spinlock for ordered_list */
+       spinlock_t list_lock;
+
+       /* Thresholding related variants */
+       atomic_t pending;
+       int max_active;
+       int current_max;
+       int thresh;
+       unsigned int count;
+       spinlock_t thres_lock;
+};
+
+struct btrfs_workqueue_struct {
+       struct __btrfs_workqueue_struct *normal;
+       struct __btrfs_workqueue_struct *high;
+};
+
+static inline struct __btrfs_workqueue_struct
+*__btrfs_alloc_workqueue(char *name, int flags, int max_active, int thresh)
+{
+       struct __btrfs_workqueue_struct *ret = kzalloc(sizeof(*ret), GFP_NOFS);
+
+       if (unlikely(!ret))
+               return NULL;
+
+       ret->max_active = max_active;
+       atomic_set(&ret->pending, 0);
+       if (thresh == 0)
+               thresh = DFT_THRESHOLD;
+       /* For low threshold, disabling threshold is a better choice */
+       if (thresh < DFT_THRESHOLD) {
+               ret->current_max = max_active;
+               ret->thresh = NO_THRESHOLD;
+       } else {
+               ret->current_max = 1;
+               ret->thresh = thresh;
+       }
+
+       if (flags & WQ_HIGHPRI)
+               ret->normal_wq = alloc_workqueue("%s-%s-high", flags,
+                                                ret->max_active,
+                                                "btrfs", name);
+       else
+               ret->normal_wq = alloc_workqueue("%s-%s", flags,
+                                                ret->max_active, "btrfs",
+                                                name);
+       if (unlikely(!ret->normal_wq)) {
+               kfree(ret);
+               return NULL;
+       }
+
+       INIT_LIST_HEAD(&ret->ordered_list);
+       spin_lock_init(&ret->list_lock);
+       spin_lock_init(&ret->thres_lock);
+       return ret;
+}
+
+static inline void
+__btrfs_destroy_workqueue(struct __btrfs_workqueue_struct *wq);
+
+struct btrfs_workqueue_struct *btrfs_alloc_workqueue(char *name,
+                                                    int flags,
+                                                    int max_active,
+                                                    int thresh)
+{
+       struct btrfs_workqueue_struct *ret = kzalloc(sizeof(*ret), GFP_NOFS);
+
+       if (unlikely(!ret))
+               return NULL;
+
+       ret->normal = __btrfs_alloc_workqueue(name, flags & ~WQ_HIGHPRI,
+                                             max_active, thresh);
+       if (unlikely(!ret->normal)) {
+               kfree(ret);
+               return NULL;
+       }
+
+       if (flags & WQ_HIGHPRI) {
+               ret->high = __btrfs_alloc_workqueue(name, flags, max_active,
+                                                   thresh);
+               if (unlikely(!ret->high)) {
+                       __btrfs_destroy_workqueue(ret->normal);
+                       kfree(ret);
+                       return NULL;
+               }
+       }
+       return ret;
+}
+
+/*
+ * Hook for threshold which will be called in btrfs_queue_work.
+ * This hook WILL be called in IRQ handler context,
+ * so workqueue_set_max_active MUST NOT be called in this hook
+ */
+static inline void thresh_queue_hook(struct __btrfs_workqueue_struct *wq)
+{
+       if (wq->thresh == NO_THRESHOLD)
+               return;
+       atomic_inc(&wq->pending);
+}
+
+/*
+ * Hook for threshold which will be called before executing the work,
+ * This hook is called in kthread content.
+ * So workqueue_set_max_active is called here.
+ */
+static inline void thresh_exec_hook(struct __btrfs_workqueue_struct *wq)
+{
+       int new_max_active;
+       long pending;
+       int need_change = 0;
+
+       if (wq->thresh == NO_THRESHOLD)
+               return;
+
+       atomic_dec(&wq->pending);
+       spin_lock(&wq->thres_lock);
+       /*
+        * Use wq->count to limit the calling frequency of
+        * workqueue_set_max_active.
+        */
+       wq->count++;
+       wq->count %= (wq->thresh / 4);
+       if (!wq->count)
+               goto  out;
+       new_max_active = wq->current_max;
+
+       /*
+        * pending may be changed later, but it's OK since we really
+        * don't need it so accurate to calculate new_max_active.
+        */
+       pending = atomic_read(&wq->pending);
+       if (pending > wq->thresh)
+               new_max_active++;
+       if (pending < wq->thresh / 2)
+               new_max_active--;
+       new_max_active = clamp_val(new_max_active, 1, wq->max_active);
+       if (new_max_active != wq->current_max)  {
+               need_change = 1;
+               wq->current_max = new_max_active;
+       }
+out:
+       spin_unlock(&wq->thres_lock);
+
+       if (need_change) {
+               workqueue_set_max_active(wq->normal_wq, wq->current_max);
+       }
+}
+
+static void run_ordered_work(struct __btrfs_workqueue_struct *wq)
+{
+       struct list_head *list = &wq->ordered_list;
+       struct btrfs_work_struct *work;
+       spinlock_t *lock = &wq->list_lock;
+       unsigned long flags;
+
+       while (1) {
+               spin_lock_irqsave(lock, flags);
+               if (list_empty(list))
+                       break;
+               work = list_entry(list->next, struct btrfs_work_struct,
+                                 ordered_list);
+               if (!test_bit(WORK_DONE_BIT, &work->flags))
+                       break;
+
+               /*
+                * we are going to call the ordered done function, but
+                * we leave the work item on the list as a barrier so
+                * that later work items that are done don't have their
+                * functions called before this one returns
+                */
+               if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
+                       break;
+               spin_unlock_irqrestore(lock, flags);
+               work->ordered_func(work);
+
+               /* now take the lock again and drop our item from the list */
+               spin_lock_irqsave(lock, flags);
+               list_del(&work->ordered_list);
+               spin_unlock_irqrestore(lock, flags);
+
+               /*
+                * we don't want to call the ordered free functions
+                * with the lock held though
+                */
+               work->ordered_free(work);
+       }
+       spin_unlock_irqrestore(lock, flags);
+}
+
+static void normal_work_helper(struct work_struct *arg)
+{
+       struct btrfs_work_struct *work;
+       struct __btrfs_workqueue_struct *wq;
+       int need_order = 0;
+
+       work = container_of(arg, struct btrfs_work_struct, normal_work);
+       /*
+        * We should not touch things inside work in the following cases:
+        * 1) after work->func() if it has no ordered_free
+        *    Since the struct is freed in work->func().
+        * 2) after setting WORK_DONE_BIT
+        *    The work may be freed in other threads almost instantly.
+        * So we save the needed things here.
+        */
+       if (work->ordered_func)
+               need_order = 1;
+       wq = work->wq;
+
+       thresh_exec_hook(wq);
+       work->func(work);
+       if (need_order) {
+               set_bit(WORK_DONE_BIT, &work->flags);
+               run_ordered_work(wq);
+       }
+}
+
+void btrfs_init_work(struct btrfs_work_struct *work,
+                    void (*func)(struct btrfs_work_struct *),
+                    void (*ordered_func)(struct btrfs_work_struct *),
+                    void (*ordered_free)(struct btrfs_work_struct *))
+{
+       work->func = func;
+       work->ordered_func = ordered_func;
+       work->ordered_free = ordered_free;
+       INIT_WORK(&work->normal_work, normal_work_helper);
+       INIT_LIST_HEAD(&work->ordered_list);
+       work->flags = 0;
+}
+
+static inline void __btrfs_queue_work(struct __btrfs_workqueue_struct *wq,
+                                     struct btrfs_work_struct *work)
+{
+       unsigned long flags;
+
+       work->wq = wq;
+       thresh_queue_hook(wq);
+       if (work->ordered_func) {
+               spin_lock_irqsave(&wq->list_lock, flags);
+               list_add_tail(&work->ordered_list, &wq->ordered_list);
+               spin_unlock_irqrestore(&wq->list_lock, flags);
+       }
+       queue_work(wq->normal_wq, &work->normal_work);
+}
+
+void btrfs_queue_work(struct btrfs_workqueue_struct *wq,
+                     struct btrfs_work_struct *work)
+{
+       struct __btrfs_workqueue_struct *dest_wq;
+
+       if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
+               dest_wq = wq->high;
+       else
+               dest_wq = wq->normal;
+       __btrfs_queue_work(dest_wq, work);
+}
+
+static inline void
+__btrfs_destroy_workqueue(struct __btrfs_workqueue_struct *wq)
+{
+       destroy_workqueue(wq->normal_wq);
+       kfree(wq);
+}
+
+void btrfs_destroy_workqueue(struct btrfs_workqueue_struct *wq)
+{
+       if (!wq)
+               return;
+       if (wq->high)
+               __btrfs_destroy_workqueue(wq->high);
+       __btrfs_destroy_workqueue(wq->normal);
+}
+
+void btrfs_workqueue_set_max(struct btrfs_workqueue_struct *wq, int max)
+{
+       wq->normal->max_active = max;
+       if (wq->high)
+               wq->high->max_active = max;
+}
+
+void btrfs_set_work_high_priority(struct btrfs_work_struct *work)
+{
+       set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
+}